Trying to use the display server as the QSV device was found to be
generally wrong in beta, so instead lets save defaults from the earlier
device enumeration similar to obs-ffmpeg-vaapi which is known to work
well.
The code assumed MFXCreateSession would always succeed, but it fails on
systems with no QSV implementations, causing a crash when we call
MFXClose later. Additional success checks for the other API calls were
also added.
Setting AdaptiveCQM to ON when LowPower is OFF crashes the
intel-oneapi-gpu implementation on Linux when using the AV1 encoder.
But using the same LowPower settings as Windows works so lets drop the
current guards.
VPL does not have a consistent initialization process for the various
backends it dispatches to. So this resolves the new VPL dispatcher
failing to initialize the video session when it dispatches to
intel-mediaSDK backend. On the oneVPL-intel-gpu backend VA-API sessions
are correctly initialized so this wasnt noticed when tested on newer
hardware.
MFXEnumImplementations expects the adapter index to be within a range of
valid Intel adapters produced by MFXLoad and the config filter. If a
system has one non-Intel high performance GPU and one Intel iGPU, OBS
and the test app will see the Intel iGPU as index 1 due to high
performance hints, but MFXEnumImplementations will expect only one valid
index, 0. In this scenario, passing a value of 1 to
MFXEnumImplementations will cause it to abort and return
MFX_ERR_NOT_FOUND (Provided index is out of possible range). This causes
subsequent capabilities testing to fail.
To avoid this, let's track how many non-Intel adapters we see and
subtract that number from adapterIdx to only pass valid index to
MFXEnumImplementations.
Enable Low Power Look-ahead (VDENC LA) after Intel Arc Graphics. Change
LA_DEPTH defaults (60 frames for normal latency and 30 frames for low
latency). Keep VME Look-ahead only for certain Intel platforms (ICL,
TGL, ADL, and RPL). Disable VME LA for HEVC
* Change GopPicSize default to 240
* Remove GopOptFlag
* Change keyint_sec default from 3 to 0
* Reduce keyint_sec minimum value from 1 to 0 (for auto)
This moves the existing adapter checking into the platform layer and
moves the Windows implementation to its platform implementation and adds
a Linux implementation based on directly querying VA-API.
Unlike Windows, this check takes ~1ms so we have no need to spin out
another thread to perform the work. This also fixes up some of the CPP/C
mixing going on in common_utils."h" to allow us to call common functions
from C files.
This cleans up the fairly arbitrary Windows includes and types and
unused functions from the common code and replaces them with platform
independent equivalents, or platform specific implementations in
common_utils.
This removes the Windows 8 support to avoid adding an additional
platform function to query this. OBS only supports Windows 10
officially, so it's about time we removed it.
Replace the checks based on D3D11 and D3D9 to check a more informative
variable aptly named UseTexAlloc for which memory allocator should be
used within the encoder.
The query being done will ovewrite ALL invalid parameters. This results
in invalid parameters that are not LowPower settings being overwritten
with 0 either introducing different errors or possibly different
behavior.
If P216 or P416 color formats are selected with QSV, these color formats
were not explicitly handled, so the switch statements would end up in
the default case. If the user had also selected a Rec. 2100 color space,
this would result in the strange error message:
"OBS does not support 8-bit output of Rec. 2100."
This message is confusing and does not correctly reflect the chosen
settings. Let's explicitly handle the P216/P416 cases and provide a more
accurate error message.
On systems with multiple graphics adapters, one card can be configured
as power saving, and another card can be configured as performance.
Sometimes, OBS and the encoder test subprocesses will not be configured
the same way, so it's necessary to provide adapter order to the encoder
test subprocesses.
This change ensures the adapter order by passing the LUIDs to the test
subprocesses. The adapter indexes will then be updated accordingly.
When this detection helper was originally written, it incorrectly set
the hardware implementation value. This fixes it to use the correct one
based upon the index.
Neither the mfxExtMasteringDisplayColourVolumme and
mfxExtContentLightLevelInfo structures appear to be supported when using
AV1, so if using AV1, do not include these structures.
It's unknown as to why AV1 fails to initialize when using either of
these extended structures as part of its configuration parameters.
Either they're completely unsupported or there's a member variable that
causes initialization to fail.